Immutable model versions
Registered artifacts have SHA-256 identities, verified before execution, with local or S3-compatible storage behind one contract.
ML deployment infrastructureA model registry and deployment control plane with immutable artifacts, weighted canaries, promotion, rollback, and runtime boundaries.

Register · canary · promote or recover
ModelForge is a model registry, deployment control plane and inference platform. Browser, CLI and API clients use the same operations to register immutable model versions, select a runtime, send a weighted share of requests to a canary and promote, abort or roll back a release.
Registered artifacts have SHA-256 identities, verified before execution, with local or S3-compatible storage behind one contract.
Weighted canaries receive controlled traffic. Promotion, abort and rollback preserve validated lifecycle transitions and environment targets.
Built-in, PyTorch, and ONNX runtimes share an interface. A standalone Go inference runtime uses the external HTTP execution contract.
Health inspection, bounded retries, circuit breakers, metrics and stable fallback govern execution at the runtime boundary.
Shipping a model is a release-management problem as well as an inference problem. An artifact must retain a verifiable identity, a candidate needs bounded exposure, and a failed new version must preserve the stable target.
Create a workspace-scoped model and immutable version, then upload and verify its artifact.

Register · canary · promote or recover
The Python/FastAPI control plane owns identity, state, traffic and tenancy. Runtimes own framework-specific loading and prediction, so a new framework does not require serving-path branches.
A canary does not become stable until promotion succeeds. Promotion and rollback update release state and environment targets transactionally.
Models, artifacts, deployments, canaries and keys are workspace-scoped. Browser sessions use OIDC/PKCE; automation uses hashed workspace API keys.
The recorded CI checkpoint passed 164 Python tests with 3 skips. A hosted CI smoke benchmark completed 200/200 requests at 133.5 requests/s. The deployment validation scenario serves a baseline, observes weighted canary traffic, promotes v2, restores v1, detects an intentionally regressed v3, and aborts it while v1 remains authoritative.
v1 remains authoritative after the deliberately failed canary.
The hosted CI smoke benchmark measures a bounded validation workload, not production capacity. ModelForge is not publicly hosted. The console preview uses illustrative example data. AWS/Terraform configuration is a deployment reference, and model caches and runtime client state are currently process-local.
A concise engineering brief and direct links to the implementation and supporting evidence.
A two-page project brief covering the purpose, workflow, design decisions, recorded checks, and limits.